Virtual laser radar system and method based on mutual warning of wind turbines

Abstract

The present invention provides a virtual laser radar system and method based on mutual early warning of wind turbines, including: a wind farm deployment data module, which obtains the front and back relationship of wind turbines arranged at intervals in the main wind direction in the wind farm; a wind speed transfer characteristic module, which obtains the transfer characteristic of the wind speed in the main wind direction in the area where the wind farm is located; a wind farm operation data module, which obtains the actual operation data of the wind farm; and a neural network module, which captures the wind speed relationship between wind turbines according to the front and back relationship of wind turbines arranged at intervals in the main wind direction in the wind farm, the transfer characteristic of the wind speed in the main wind direction, and the wind farm operation data, so as to replace the physical laser radar to predict gusts.

Description

Technical Field

[0001] The present invention relates to the technical field of wind turbines, and in particular to a virtual laser radar system and method based on mutual early warning of wind turbines. Background Art

[0002] With the increasing tension in energy supply and demand and the increasing concern of human beings about environmental climate change, many European and American countries have raised the development of renewable clean energy to a national strategic level. As a representative of clean, environmentally friendly and renewable energy, wind energy has a large content, a wide distribution range and good development prospects. Developed countries in Europe and the United States have invested huge resources in the research and application of wind power technology.

[0003] In the development of wind power technology, wind speed prediction has been a hot topic in recent years. Wind speed is affected by many factors such as temperature, air pressure, and terrain, and has a strong randomness. Wind speed prediction can be divided into long-term (months to years), medium-term (days to weeks), short-term (hours), and ultra-short-term (minutes) predictions according to the prediction time scale. These wind speed predictions mainly serve the wind power prediction of wind farms, so as to formulate correct deployment plans and reduce the adverse effects of wind power on the power grid.

[0004] Although there are many wind speed prediction methods at home and abroad, most of them are concentrated on predictions on time scales above the minute level, such as Kalman filtering method, time series analysis method, neural network method and fuzzy logic method. These wind speed prediction methods basically build a model for the original wind speed time series of the wind farm, and then predict the future wind speed, or convert the non-stationary wind speed series into a stationary series by means of difference, and then model and predict the stationary series. However, for second-level wind speed prediction, relatively reliable prediction methods are rare. In "The Impact of Control Technology—2nd Edition" published by IEEE, advance planning and control based on future second-level wind speed time series were proposed as a major challenge for future wind turbine control research.

[0005] In recent years, a method of using LiDAR to predict wind speed in seconds has emerged. LiDAR emits lasers forward to measure the wind speed in three dimensions a few hundred meters in front of the wind turbine, thereby predicting the wind speed in the next few seconds in advance. However, due to its high price, being greatly affected by weather, instability, and difficulty in calibration, the actual effect is not good.

[0006] In summary, physical lidar has poor performance in actual use in wind farms and is difficult to apply. In addition, (1) physical lidar is expensive; (2) physical lidar is greatly affected by weather; (3) physical lidar can only predict wind speed within a short time, about 10 seconds; and (4) physical lidar calibration is difficult to set up. Summary of the invention

[0007] The purpose of the present invention is to provide a virtual laser radar system and method based on mutual warning of wind turbines, so as to solve the problem that the existing physical laser radar has poor effect in actual use in wind farms.

[0008] In order to solve the above technical problems, the present invention provides a virtual laser radar system based on wind turbine mutual warning, comprising:

[0009] The wind farm deployment data module obtains the front-to-back relationship of wind turbines arranged at intervals in the main wind direction in the wind farm; due to the poor performance of hardware laser radar, the present invention uses a software method to replace the laser radar to predict gusts. Because wind speed is transitive, wind farms are deployed with front-to-back relationships in the main wind direction. If the front row of wind turbines encounters a gust, the rear row of wind turbines will also encounter the same gust after a period of time. Based on this physical principle and observation of wind farm operation data, it can be seen that the wind speeds between wind turbines are related and have a certain lag characteristic. The present invention uses a neural network to capture this relationship.

[0010] A wind speed transfer characteristic module, which obtains the transfer characteristics of wind speed in the main wind direction in the area where the wind farm is located; a wind farm operation data module, which obtains the actual operation data of the wind farm; and

[0011] The neural network module captures the wind speed relationship between wind turbines based on the front-to-back relationship of wind turbines arranged at intervals in the main wind direction within the wind farm, the transmission characteristics of wind speed in the main wind direction, and wind farm operation data, to replace the physical lidar to predict gusts.

[0012] Optionally, in the virtual laser radar system based on wind turbine mutual warning, the neural network module includes a recurrent neural network model and a graph neural network model;

[0013] The recurrent neural network model and the graph neural network model are formed based on the front-to-back relationship of wind turbines arranged at intervals in the main wind direction in the wind farm, and the transmission characteristics of wind speed in the main wind direction;

[0014] The neural network module uses the actual operation data of the wind farm to train the recurrent neural network model and the graph neural network model to perform spatiotemporal prediction of wind speed, and uses the actual operation data of the wind farm in other time periods for verification. Based on the recurrent neural network and the graph neural network, the model is trained using the actual operation data of the wind farm for spatiotemporal prediction, and is verified using data from other time periods. The results show that the practicality of the present invention exceeds that of physical laser radar, and it can effectively perform second-level wind speed prediction in actual wind farms, which has great economic value in reducing wind turbine fatigue and increasing power generation.

[0015] Optionally, in the virtual laser radar system based on mutual early warning of wind turbines, the actual operation data of the wind farm is second-level wind speed detection data;

[0016] The neural network module captures the wind speed relationship between wind turbines by: obtaining the correlation of the second-level wind speed detection data of each wind turbine based on the second-level wind speed detection data, and obtaining the flow field information of the entire wind farm based on the correlation of the second-level wind speed detection data of each wind turbine.

[0017] Optionally, in the virtual laser radar system based on mutual warning of wind turbines, the transmission characteristic of the wind speed in the main wind direction is a mutual warning characteristic between wind turbines;

[0018] The neural network module combines the front and back relationship of wind turbines arranged at intervals in the main wind direction in the wind farm, uses the time series deep learning algorithm to train the second-level wind speed detection data, establishes a specific wind field flow field model, and predicts the wind speed trend information of each wind turbine in the next tens of seconds in real time based on the specific wind field flow field model, and uses the mutual warning characteristics between wind turbines to realize virtual lidar.

[0019] Each wind farm has a huge amount of second-level wind speed data. The wind speed data of different positions are not isolated and unrelated. In these massive data, their mutual relationships have described the flow field information of the entire wind farm. Therefore, combined with the position arrangement data of the wind farm, the time series deep learning algorithm can be used to train these wind speed data to establish a flow field model for a specific wind farm, and based on the flow field model, the wind speed trend information of each wind turbine in the next few tens of seconds can be predicted in real time, and then the mutual warning between wind turbines can be used to realize virtual lidar.

[0020] Optionally, in the virtual laser radar system based on mutual warning of wind turbines, the spatiotemporal prediction of wind speed includes: in a specific technical field, in the traditional time series prediction problem, the historical data itself is often used for learning to find out the law, that is, the relationship between the time series before and after is captured by a recurrent neural network. For example, a structural gate regression unit commonly used in recurrent neural networks.

[0021] The wind speed historical data of this wind turbine and other wind turbines are used for deep learning, and the recurrent neural network gate regression unit is used to capture the relationship between the time series, including:

[0022] z t =σ(W z x t +U z h t-1 )

[0023] r t =σ(W t x t +U t h t-1 )

[0024]

[0025]

[0026] Z t It is the update gate, which indicates whether the information at the previous moment needs to be updated;

[0027] r t It is the reset gate, which indicates whether the information at the previous moment needs to be reset;

[0028] is a candidate output, receiving x t 、h t-1 Time information;

[0029] h t is the hidden layer output of the gate regression unit, receiving h t-1 、h t Time information;

[0030] The previous and next relationships in the time series are captured through the training of the gate regression unit, and the future wind speed data is predicted through the historical wind speed data of this wind turbine and other wind turbines.

[0031] Optionally, in the virtual laser radar system based on mutual warning of wind turbines, performing spatiotemporal prediction of wind speed also includes: in terms of spatial structure problem, a method of graph neural network GNN has emerged in the past two years, that is, for a graph G = (V, E), V is an edge set in the graph, E is a node set in the graph, and the attributes of each node in E are affected by the attributes of the connected nodes;

[0032] X is the original node attribute, A is the adjacency matrix, Z is the transformed node attribute, I is a diagonal matrix, and for a two-layer graph neural network, the activation functions are ReLU and Softmax respectively; then

[0033]

[0034]

[0035] Graph neural networks are used to capture the relationships between spatial nodes and obtain the patterns of data flow in the relationships between spatial nodes. Currently, graph neural networks have been used in traffic flow prediction, social networks, protein structure prediction, etc. No one has used them in the field of wind speed prediction, let alone in wind farms.

[0036] Optionally, in the virtual laser radar system based on mutual warning of wind turbines, not only the recurrent neural network is used to capture the relationship in the time series of a single wind turbine, but also the graph neural network is used to capture the spatial structural relationship between multiple wind turbines in the wind farm. Not only the historical data of the wind turbine itself is used, but also the historical data of other wind turbines in the wind farm are used. By observing the wind speed time series of the adjacent wind turbines, the obvious lag is obtained, that is, the wind speed of the upwind wind turbine at the current time t will also appear at the downwind wind turbine at time t+Δt, especially in the case of gusts, where Δt depends on the distance d between the two wind turbines and the current wind speed ws,

[0037] The wind speed of a certain wind turbine at time t+Δt is predicted by the wind speed of the remaining wind turbines at time t and the relative position of the wind turbines. In a certain wind direction, excluding the wind speed attenuation, the wind speed of the first wind turbine at time t is equal to the wind speed of the second wind turbine at time t+Δt, Δt≈d / ws, where d is the distance between the two wind turbines and ws is the wind speed of the first wind turbine at time t. The above method provides a solid foundation for predicting gusts in advance for downwind wind turbines. At the same time, all wind turbines in the entire wind farm form a network, and the future wind speed of a certain wind turbine can be effectively predicted through the historical wind speed time series and relative position relationship of the remaining wind turbines.

[0038] Optionally, in the virtual laser radar system based on wind turbine mutual warning, wind farm time series data is collected, data cleaning and standardization are performed, training data sets and test data sets are divided, wind farm network structure is constructed, deep learning neural network model is trained, model verification, model deployment, and wind turbine control are performed. Specifically, the following are performed: wind speed data of each wind turbine in the wind farm is collected;

[0039] The actual operating data of wind speed is aligned by time, and the abnormal data is cleaned according to the operating status of the wind turbine to standardize the data. The mean is subtracted and divided by the standard deviation to ensure that the mean of the standardized data is 0 and the variance is 1:

[0040]

[0041] The training set and the test set are divided according to the time sequence, where the data of the training set accounts for 80% of the actual wind speed operation data, and the data of the test set accounts for 20% of the actual wind speed operation data. The training set is used to train the network weights, and the test set does not participate in any training, but is only used to calculate the evaluation index to detect the model performance;

[0042] Constructing a wind farm network structure, wherein the wind farm network structure is divided into a static network and a dynamic network, wherein the static network is a relationship matrix between fixed wind turbines during the training process, and the dynamic network is a relationship matrix between unfixed wind turbines during the training process, and the relationship matrix changes with the similarity of the current wind speed time series of different wind turbines;

[0043] The deep learning neural network model is trained. The graph convolutional neural network inputs the standardized historical wind speed data of each wind turbine and the adjacency matrix to extract the spatial feature matrix. The output result is input into the recurrent neural network model to extract the temporal feature matrix and output the wind speed value of each wind turbine at the future moment. The optimizer is determined to be Adam, and the loss function is the root mean square error (MSE).

[0044] After the model training set converges, it is verified on the test set, the mean absolute error Mae index is calculated and the time series is plotted, and the model is deployed at the wind farm station;

[0045] Using virtual LiDAR for wind turbine control.

[0046] The "virtual laser radar" method can effectively predict the wind speed of each wind turbine in the wind farm for a period of time in the future. In principle, the wind turbines in the front row in the incoming wind direction mainly rely on their own historical data for prediction. In addition to their own historical laws, the wind turbines in the back row can also rely on the information provided by the front row of wind turbines for prediction, which is more accurate, especially when all wind turbines are operating normally, so as to effectively replace the laser radar.

[0047] Optionally, in the virtual laser radar system based on mutual warning of wind turbines, the wind speed trend prediction time length based on the mutual warning model of wind turbines is much longer than the predictable range of hardware laser radar. The wind speed measurement range of the usual nacelle laser radar is about 80 to 200 meters in front of the wind rotor, while the wind speed prediction range based on deep learning can reach more than 300 meters. For wind turbines, the longer the wind speed time series prediction range, the more suitable it is for it to make future control behavior decisions as a single intelligent agent. The virtual laser radar based on mutual warning of wind turbines is based on the wind speed prediction time series of deep learning, and is equipped with a model predictive control algorithm and a Markov decision model to calculate the optimal power generation and component load of the wind turbine.

[0048] The present invention also provides a virtual laser radar method based on wind turbine mutual warning, comprising:

[0049] The wind farm deployment data module obtains the front-to-back relationship of wind turbines arranged at intervals in the main wind direction in the wind farm;

[0050] The wind speed transfer characteristic module obtains the transfer characteristics of the wind speed in the main wind direction in the area where the wind farm is located;

[0051] The wind farm operation data module obtains actual operation data of the wind farm; and

[0052] The neural network module captures the wind speed relationship between wind turbines based on the front-to-back relationship of wind turbines arranged at intervals in the main wind direction in the wind farm, the transmission characteristics of wind speed in the main wind direction, and the wind farm operation data, to replace the physical lidar to predict gusts.

[0053] In the virtual laser radar system and method based on mutual warning of wind turbines provided by the present invention, the wind speed relationship between wind turbines is captured through a neural network module based on the front and rear relationship of wind turbines arranged at intervals in the main wind direction in the wind farm, the transmission characteristics of wind speed in the main wind direction, and wind farm operation data, so as to replace the physical laser radar to predict gusts, thereby realizing a virtual laser radar method based on mutual warning of wind turbines. This method is more practical than laser radar, and can effectively perform second-level wind speed prediction in actual wind farms, which has great economic value in reducing wind turbine fatigue and increasing power generation.

[0054] The present invention proposes a virtual laser radar system and method based on mutual warning of wind turbines, which predicts the wind speed at each wind turbine site in the future through deep learning of historical wind speed data. At the same time, the predicted second-level wind speed is applied to the real-time control of wind turbines to overcome the various limitations brought by hardware laser radar.

[0055] The beneficial effects of the present invention are: cost saving, which can save about 500,000 yuan in costs for each laser radar; it is not affected by weather, has high reliability, accuracy, and stable application; it can obtain more comprehensive and detailed wind conditions with a longer span; the software architecture is easy to deploy, does not require high computing resources, and has no deployment threshold; based on second-level wind speed time series prediction, it can protect the wind turbine from damage by extreme loads and effectively reduce tower fatigue loads, which is also of great significance to reducing the cost of future wind turbine design. BRIEF DESCRIPTION OF THE DRAWINGS

[0056] Figure 1 It is a schematic diagram of a virtual laser radar method based on mutual early warning of wind turbines according to an embodiment of the present invention;

[0057] Figure 2 It is a schematic diagram of a gate regression unit of a virtual laser radar system based on mutual warning of wind turbines according to an embodiment of the present invention;

[0058] Figure 3 It is a schematic diagram of a neural network of a virtual laser radar system diagram based on mutual warning of wind turbines according to an embodiment of the present invention;

[0059] Figure 4It is a schematic diagram of a virtual laser radar method based on mutual warning of wind turbines according to an embodiment of the present invention. DETAILED DESCRIPTION

[0060] The virtual laser radar system and method based on wind turbine mutual warning proposed by the present invention are further described in detail below in combination with the accompanying drawings and specific embodiments. The advantages and features of the present invention will become clearer according to the following description and claims. It should be noted that the drawings are all in a very simplified form and are not in precise proportions, which are only used to conveniently and clearly assist in explaining the purpose of the embodiments of the present invention.

[0061] In addition, unless otherwise specified, the features in different embodiments of the present invention may be combined with each other. For example, a feature in the second embodiment may be used to replace a corresponding feature in the first embodiment or a feature with the same or similar function, and the resulting embodiment also falls within the disclosure or description scope of the present application.

[0062] The core idea of ​​the present invention is to provide a virtual laser radar system and method based on mutual warning of wind turbines to solve the problem that the existing physical laser radar has poor effect in actual use in wind farms.

[0063] To realize the above idea, the present invention provides a virtual lidar system and method based on mutual warning of wind turbines, including: a wind farm deployment data module, which obtains the front and back relationship of wind turbines arranged at intervals in the main wind direction in the wind farm; a wind speed transfer characteristic module, which obtains the transfer characteristics of the wind speed in the main wind direction in the area where the wind farm is located; a wind farm operation data module, which obtains the actual operation data of the wind farm; and a neural network module, which captures the wind speed relationship between wind turbines based on the front and back relationship of wind turbines arranged at intervals in the main wind direction in the wind farm, the transfer characteristics of the wind speed in the main wind direction, and the wind farm operation data, so as to replace the physical lidar to predict gusts.

[0064] Since the use of hardware laser radar is not effective, the present invention uses a software method to replace the laser radar to predict gusts. Because wind speed is transferable, the wind farm will be arranged in a front-to-back relationship in the main wind direction. If the front row of fans encounters a gust, the rear row of fans will also encounter the same gust after a period of time. Based on this physical principle and the observation of wind farm operation data, it can be seen that the wind speeds between fans are related and have a certain lag characteristic. The present invention uses a neural network to capture this relationship.

[0065] This embodiment provides a virtual laser radar system based on mutual warning of wind turbines, such as Figure 1As shown, it includes: a wind farm deployment data module, which obtains the front and back relationship of wind turbines arranged at intervals in the main wind direction in the wind farm; a wind speed transfer characteristic module, which obtains the transfer characteristics of wind speed in the main wind direction in the area where the wind farm is located; a wind farm operation data module, which obtains the actual operation data of the wind farm; and a neural network module, which captures the wind speed relationship between wind turbines according to the front and back relationship of wind turbines arranged at intervals in the main wind direction in the wind farm, the transfer characteristics of wind speed in the main wind direction, and the wind farm operation data, so as to replace the physical laser radar to predict gusts.

[0066] Specifically, in the virtual laser radar system based on mutual warning of wind turbines, the neural network module includes a recurrent neural network model and a graph neural network model; the recurrent neural network model and the graph neural network model are formed according to the front-to-back relationship of wind turbines arranged at intervals in the main wind direction in the wind farm, and the transmission characteristics of wind speed in the main wind direction; the neural network module uses the actual operation data of the wind farm to train the recurrent neural network model and the graph neural network model to perform spatiotemporal prediction of wind speed, and uses the actual operation data of the wind farm in other time periods for verification. Based on the recurrent neural network and the graph neural network, the model is trained using the actual operation data of the wind farm for spatiotemporal prediction, and is verified using data from other time periods. The results show that the practicality of the present invention exceeds that of physical laser radar, and it can effectively perform second-level wind speed prediction in actual wind farms, which has great economic value in reducing wind turbine fatigue and increasing power generation.

[0067] Furthermore, in the virtual lidar system based on mutual warning of wind turbines, the actual operation data of the wind farm is second-level wind speed detection data; the neural network module captures the wind speed relationship between wind turbines, including: according to the second-level wind speed detection data, obtaining the correlation of the second-level wind speed detection data of each wind turbine, and according to the correlation of the second-level wind speed detection data of each wind turbine, obtaining the flow field information of the entire wind farm.

[0068] In addition, in the virtual laser radar system based on mutual warning of wind turbines, the transmission characteristics of the wind speed in the main wind direction are the mutual warning characteristics between wind turbines; the neural network module combines the front and back relationship of wind turbines arranged at intervals in the main wind direction in the wind farm, and uses the time series deep learning algorithm to train the second-level wind speed detection data, establishes a specific wind field flow field model, and predicts the wind speed trend information of each wind turbine in the next tens of seconds in real time based on the specific wind field flow field model, and uses the mutual warning characteristics between wind turbines to realize the virtual laser radar. Each wind farm has a large amount of second-level wind speed data, and the wind speed data of different positions are not isolated and unrelated. In these massive data, their mutual relationship has described the flow field information of the entire wind farm. Therefore, combined with the position arrangement data of the wind farm, the time series deep learning algorithm can be used to train these wind speed data, so as to establish a flow field model of a specific wind farm, and predict the wind speed trend information of each wind turbine in the next tens of seconds in real time based on the flow field model, and then use the mutual warning between wind turbines to realize the virtual laser radar.

[0069] In one embodiment of the present invention, in the virtual laser radar system based on wind turbine mutual warning, the spatiotemporal prediction of wind speed includes: in the specific technical field, in the traditional time series prediction problem, the historical data itself is often used for learning to find the law, that is, the recurrent neural network is used to capture the relationship between the time series before and after. For example, a structural gate regression unit commonly used in the recurrent neural network, such as Figure 2 Using the historical wind speed data of this wind turbine and other wind turbines for deep learning, the recurrent neural network gate regression unit is used to capture the relationship between the time series, including:

[0070] z t =σ(W z x t +U z h t-1 )

[0071] r t =σ(W t x t +U t h t-1 )

[0072]

[0073]

[0074] Z t It is the update gate, which indicates whether the information at the previous moment needs to be updated;

[0075] r t It is the reset gate, which indicates whether the information at the previous moment needs to be reset;

[0076] is a candidate output, receiving x t 、h t-1 Time information;

[0077] h t is the hidden layer output of the gate regression unit, receiving h t-1 、h t Time information;

[0078] The previous and next relationships in the time series are captured through the training of the gate regression unit, and the future wind speed data is predicted through the historical wind speed data of this wind turbine and other wind turbines.

[0079] In another embodiment of the present invention, in the virtual laser radar system based on wind turbine mutual warning, performing spatiotemporal wind speed prediction also includes: In terms of spatial structure problems, graph neural network GNN methods have emerged in the past two years, such as Figure 3 As shown, for a graph G = (V, E), V is the edge set in the graph, E is the node set in the graph, and the attributes of each node in E are affected by the attributes of the connected nodes; X is the original node attribute, A is the adjacency matrix, Z is the transformed node attribute, and I is a diagonal matrix. For a two-layer graph neural network, the activation functions are ReLU and Softmax respectively; then

[0080]

[0081]

[0082] Graph neural networks are used to capture the relationships between spatial nodes and obtain the patterns of data flow in the relationships between spatial nodes. Currently, graph neural networks have been used in traffic flow prediction, social networks, protein structure prediction, etc. No one has used them in the field of wind speed prediction, let alone in wind farms.

[0083] Specifically, in the virtual laser radar system based on mutual warning of wind turbines, not only the recurrent neural network is used to capture the relationship in the time series of a single wind turbine, but also the graph neural network is used to capture the spatial structural relationship between multiple wind turbines in the wind farm. Not only the historical data of the wind turbine itself is used, but also the historical data of other wind turbines in the wind farm are used. By observing the wind speed time series of the adjacent wind turbines, an obvious lag is obtained, that is, the wind speed of the upwind wind turbine at the current time t will also appear at the downwind wind turbine at time t+Δt, especially in the case of gusts, where Δt depends on the distance d between the two wind turbines and the current wind speed ws. The wind speed of a certain wind turbine at time t+Δt is predicted by the wind speed of the remaining wind turbines at time t and the relative position of the wind turbines. In a certain wind direction, excluding the wind speed attenuation, the wind speed of the first wind turbine at time t is equal to the wind speed of the second wind turbine at time t+Δt, Δt≈d / ws, where d is the distance between the two wind turbines and ws is the wind speed of the first wind turbine at time t. The above method provides a solid foundation for predicting gusts in advance for downwind wind turbines. At the same time, all wind turbines in the entire wind farm form a network, and the future wind speed of a certain wind turbine can be effectively predicted through the historical wind speed time series and mutual position relationship of the remaining wind turbines.

[0084] Furthermore, in the virtual laser radar system based on wind turbine mutual warning, wind farm time series data is collected, data cleaning and standardization are performed, training data sets and test data sets are divided, wind farm network structure is constructed, deep learning neural network model is trained, model verification, model deployment, and wind turbine control are performed. Figure 4 As shown, it specifically includes: collecting the operating wind speed data of each wind turbine in the wind farm; aligning the actual operating wind speed data by time, cleaning the abnormal data according to the operating status of the wind turbine to standardize the data, and subtracting the mean and dividing by the standard deviation to ensure that the standardized data has a mean of 0 and a variance of 1:

[0085]

[0086] The training set and the test set are divided into a time sequence, wherein the data of the training set accounts for 80% of the actual operation data of the wind speed, and the data of the test set accounts for 20% of the actual operation data of the wind speed. The training set is used to train the network weights, and the test set does not participate in any training, and is only used to calculate the evaluation index to detect the model performance; a wind farm network structure is constructed, wherein the wind farm network structure is divided into a static network and a dynamic network, wherein the static network is a relationship matrix between fixed wind turbines during the training process, and the dynamic network is a relationship matrix between wind turbines that are not fixed during the training process, and the relationship matrix changes with the similarity of the current wind speed time series of different wind turbines; a deep learning neural network model is trained, wherein the graph convolutional neural network inputs the standardized historical wind speed data and adjacency matrix of each wind turbine to extract the spatial feature matrix, and the output result is input into the recurrent neural network model to extract the time feature matrix, and output the wind speed value of each wind turbine at the future moment; the optimizer is determined to be Adam, and the loss function is the root mean square error (MSE); after the model training set converges, it is verified in the test set, the mean absolute error (Mae) indicator is calculated and the time series is drawn, and the model is deployed at the wind farm station; a virtual laser radar is used for wind turbine control.

[0087] The present invention can effectively predict the wind speed of each wind turbine in the wind farm in the future through the "virtual laser radar" method. In principle, the front row wind turbines in the incoming wind direction mainly rely on their own historical data for prediction, and the rear row wind turbines can rely on the information provided by the front row wind turbines in addition to their own historical laws for prediction, which is more accurate, especially when all wind turbines are operating normally, to effectively replace the laser radar.

[0088] In addition, in the virtual laser radar system based on mutual warning of wind turbines, the wind speed trend prediction time length based on the mutual warning model of wind turbines is much longer than the predictable range of hardware laser radar. The wind speed measurement range of the usual nacelle laser radar is about 80 to 200 meters in front of the wind rotor, while the wind speed prediction range based on deep learning can reach more than 300 meters. For wind turbines, the longer the wind speed time series prediction range, the more suitable it is for it to make future control behavior decisions as a single intelligent agent. The virtual laser radar based on mutual warning of wind turbines is based on the wind speed prediction time series of deep learning, and is equipped with a model predictive control algorithm and a Markov decision model to calculate the optimal power generation and component load of the wind turbine.

[0089] In the virtual laser radar system and method based on mutual warning of wind turbines provided by the present invention, the wind speed relationship between wind turbines is captured through a neural network module based on the front and rear relationship of wind turbines arranged at intervals in the main wind direction in the wind farm, the transmission characteristics of wind speed in the main wind direction, and wind farm operation data, so as to replace the physical laser radar to predict gusts, thereby realizing a virtual laser radar method based on mutual warning of wind turbines. This method is more practical than laser radar, and can effectively perform second-level wind speed prediction in actual wind farms, which has great economic value in reducing wind turbine fatigue and increasing power generation.

[0090] The present invention proposes a virtual laser radar system and method based on mutual warning of wind turbines, which predicts the wind speed at each wind turbine site in the future through deep learning of historical wind speed data. At the same time, the predicted second-level wind speed is applied to the real-time control of wind turbines to overcome the various limitations brought by hardware laser radar.

[0091] The present invention can realize wind speed prediction without using any physical laser radar. The existing virtual laser radar is mostly based on the physical laser radar installed on the front fan to predict the wind speed ahead before realizing the function of the virtual laser radar. The virtual laser radar in the present invention does not need to install any physical laser radar. It predicts the wind speed that the front fan is about to face through the historical wind speed data, and uses the original wind speed detection equipment on the fan to detect the real-time wind speed of the fan. The two are combined to train a model, and then the model is used to predict the future wind speed of the front fan. The installation of any physical laser radar is avoided, and all the disadvantages of the physical laser radar can be avoided. The prior art only needs to use one physical laser radar, which will inevitably encounter the problems of high price, great influence by weather, short wind speed forecast time, and difficult calibration. The great significance of the present invention lies in avoiding the installation of any physical laser radar.

[0092] The beneficial effects of the present invention are: cost saving, which can save about 500,000 yuan in costs for each laser radar; it is not affected by weather, has high reliability, accuracy, and stable application; it can obtain more comprehensive and detailed wind conditions with a longer span; the software architecture is easy to deploy, does not require high computing resources, and has no deployment threshold; based on second-level wind speed time series prediction, it can protect the wind turbine from damage by extreme loads and effectively reduce tower fatigue loads, which is also of great significance to reducing the cost of future wind turbine design.

[0093] In summary, the above embodiments describe in detail different configurations of the virtual laser radar system and method based on mutual warning of wind turbines. Of course, the present invention includes but is not limited to the configurations listed in the above embodiments. Any content that is transformed based on the configurations provided in the above embodiments belongs to the scope of protection of the present invention. Those skilled in the art can draw inferences from the contents of the above embodiments.

[0094] In this specification, each embodiment is described in a progressive manner, and each embodiment focuses on the differences from other embodiments. The same or similar parts between the embodiments can be referred to each other. For the system disclosed in the embodiment, since it corresponds to the method disclosed in the embodiment, the description is relatively simple, and the relevant parts can be referred to the method part.

[0095] The above description is only a description of the preferred embodiments of the present invention, and is not intended to limit the scope of the present invention. Any changes or modifications made by a person skilled in the art in the field of the present invention based on the above disclosure shall fall within the scope of protection of the claims.

Claims

1. A virtual laser radar system based on mutual warning of wind turbines, characterized in that: include: A wind farm deployment data module is configured to obtain the front-to-back relationship of wind turbines arranged at intervals in the main wind direction in the wind farm; A wind speed transfer characteristic module is configured to obtain the transfer characteristic of the wind speed in the main wind direction in the area where the wind farm is located; A wind farm operation data module is configured to obtain actual operation data of the wind farm; as well as A neural network module is configured to capture the wind speed relationship between wind turbines according to the front-to-back relationship of wind turbines arranged at intervals in the main wind direction in the wind farm, the transfer characteristics of wind speed in the main wind direction, and wind farm operation data, so as to replace the physical laser radar to predict gusts, wherein the neural network module includes a recurrent neural network model and a graph neural network model, the recurrent neural network model and the graph neural network model are formed according to the front-to-back relationship of wind turbines arranged at intervals in the main wind direction in the wind farm, and the transfer characteristics of wind speed in the main wind direction, the neural network module uses the actual operation data of the wind farm to train the recurrent neural network model and the graph neural network model, performs spatiotemporal wind speed prediction, and uses the actual operation data of the wind farm in other time periods for verification, wherein the transfer characteristics of wind speed in the main wind direction are mutual warning characteristics between wind turbines, the neural network module combines the front-to-back relationship of wind turbines arranged at intervals in the main wind direction in the wind farm, uses the time series deep learning algorithm to train the second-level wind speed detection data, establishes a specific wind field flow field model, and predicts the wind speed trend information of each wind turbine in the next tens of seconds in real time based on the specific wind field flow field model, and realizes the virtual laser radar by using the mutual warning characteristics between wind turbines; The spatiotemporal wind speed forecast includes: The wind speed historical data of this wind turbine and other wind turbines are used for deep learning, and the recurrent neural network gate regression unit is used to capture the relationship between the time series, including: z t =σ(W z x t +U z h t-1 ) r t =σ(W t x t +U t h t-1 ) Z t It is the update gate, which indicates whether the information at the previous moment needs to be updated; r t It is the reset gate, which indicates whether the information at the previous moment needs to be reset; is a candidate output, receiving x t 、h t-1 Time information; h t is the hidden layer output of the gate regression unit, receiving h t-1 、h t Time information; The previous and next relationships in the time series are captured through the training of the gate regression unit, and the future wind speed data is predicted through the historical wind speed data of this wind turbine and other wind turbines.

2. The virtual laser radar system based on wind turbine mutual warning as claimed in claim 1 is characterized in that: The actual operation data of the wind farm is the wind speed detection data at the second level; The neural network module captures the wind speed relationship between wind turbines by: obtaining the correlation of the second-level wind speed detection data of each wind turbine based on the second-level wind speed detection data, and obtaining the flow field information of the entire wind farm based on the correlation of the second-level wind speed detection data of each wind turbine.

3. The virtual laser radar system based on wind turbine mutual warning as claimed in claim 1 is characterized in that: The spatiotemporal forecast of wind speed also includes: Capture the mutual relationship between spatial nodes through graph neural network and obtain the law of data flow in the mutual relationship between spatial nodes; G = (V, E), V is the edge set in the graph, E is the node set in the graph, and the attributes of each node in E are affected by the attributes of the connected nodes; X is the original node attribute, A is the adjacency matrix, Z is the transformed node attribute, I is a diagonal matrix, and for a two-layer graph neural network, the activation functions are ReLU and Softmax respectively; 4. The virtual laser radar system based on wind turbine mutual warning as claimed in claim 1 is characterized in that: The wind speed of a certain wind turbine at time t+Δt is predicted by the wind speed of the other wind turbines at time t and the relative position relationship of the wind turbines. In a certain wind direction, excluding wind speed attenuation, the wind speed of the first wind turbine at time t is equal to the wind speed of the second wind turbine at time t+Δt, Δt≈d / ws, where d is the distance between the two wind turbines, and ws is the wind speed of the first wind turbine at time t.

5. The virtual laser radar system based on wind turbine mutual warning as claimed in claim 1 is characterized in that: The actual operating data of wind speed is aligned by time, and the abnormal data is cleaned according to the operating status of the wind turbine to standardize the data. The mean is subtracted and divided by the standard deviation to ensure that the mean of the standardized data is 0 and the variance is 1: The training set and the test set are divided according to the time sequence, where the data of the training set accounts for 80% of the actual wind speed operation data, and the data of the test set accounts for 20% of the actual wind speed operation data. The training set is used to train the network weights, and the test set is used to calculate the evaluation index to detect the model performance; Constructing a wind farm network structure, wherein the wind farm network structure is divided into a static network and a dynamic network, wherein the static network is a relationship matrix between fixed wind turbines during the training process, and the dynamic network is a relationship matrix between unfixed wind turbines during the training process, and the relationship matrix changes with the similarity of the current wind speed time series of different wind turbines; Train the deep learning neural network model. The graph convolutional neural network inputs the standardized historical wind speed data of each wind turbine and the adjacency matrix to extract the spatial feature matrix. The output results are input into the recurrent neural network model to extract the temporal feature matrix and output the wind speed value of each wind turbine at the future time. Determine the optimizer as Adam and the loss function as root mean square error (MSE); After the model training set converges, it is verified on the test set, the mean absolute error Mae index is calculated and the time series is plotted, and the model is deployed at the wind farm station; Using virtual LiDAR for wind turbine control.

6. The virtual laser radar system based on wind turbine mutual warning as claimed in claim 1, characterized in that: The virtual lidar based on mutual warning of wind turbines is based on the wind speed prediction time series based on deep learning, and is equipped with model predictive control algorithm and Markov decision model to calculate the optimal comprehensive power generation and component load of the wind turbine.

7. A virtual laser radar method based on mutual warning of wind turbines, characterized in that: include: The wind farm deployment data module obtains the front-to-back relationship of wind turbines arranged at intervals in the main wind direction in the wind farm; The wind speed transfer characteristic module obtains the transfer characteristics of the wind speed in the main wind direction in the area where the wind farm is located; The wind farm operation data module obtains the actual operation data of the wind farm; as well as The neural network module captures the wind speed relationship between wind turbines according to the front-to-back relationship of wind turbines arranged at intervals in the main wind direction in the wind farm, the transfer characteristics of wind speed in the main wind direction, and the wind farm operation data, so as to replace the physical laser radar to predict gusts, wherein the neural network module includes a recurrent neural network model and a graph neural network model, the recurrent neural network model and the graph neural network model are formed according to the front-to-back relationship of wind turbines arranged at intervals in the main wind direction in the wind farm, and the transfer characteristics of wind speed in the main wind direction, the neural network module uses the actual operation data of the wind farm to train the recurrent neural network model and the graph neural network model, performs spatiotemporal wind speed prediction, and uses the actual operation data of the wind farm in other time periods for verification, wherein the transfer characteristics of wind speed in the main wind direction are mutual warning characteristics between wind turbines, the neural network module combines the front-to-back relationship of wind turbines arranged at intervals in the main wind direction in the wind farm, uses the time series deep learning algorithm to train the second-level wind speed detection data, establishes a specific wind field flow field model, and predicts the wind speed trend information of each wind turbine in the next tens of seconds in real time based on the specific wind field flow field model, and realizes the virtual laser radar by using the mutual warning characteristics between wind turbines; The spatiotemporal wind speed forecast includes: The wind speed historical data of this wind turbine and other wind turbines are used for deep learning, and the recurrent neural network gate regression unit is used to capture the relationship between the time series, including: z t =σ(W z x t +U z h t-1 ) r t =σ(W y x y +U y h t-1 ) Z t It is the update gate, which indicates whether the information at the previous moment needs to be updated; r t It is the reset gate, which indicates whether the information at the previous moment needs to be reset; is a candidate output, receiving x t 、h t-1 Time information; h t is the hidden layer output of the gate regression unit, receiving h t-1 、h t Time information; The previous and next relationships in the time series are captured through the training of the gate regression unit, and the future wind speed data is predicted through the historical wind speed data of this wind turbine and other wind turbines.

Images